Magnetic resonance image processing transformer for general accelerated image restoration.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning models achieve state-of-the-art performance in imaging tasks.
• Vision Transformer (ViT) architectures excel at feature extraction after large-scale pre-training.
• Accelerated Magnetic Resonance Imaging (MRI) requires robust restoration methods.

Purpose of the Study:

• Introduce the Magnetic Resonance Image Processing Transformer (MR-IPT), a ViT-based framework for accelerated MRI restoration.
• Enhance the generalizability and robustness of MRI restoration models.
• Develop a unified framework for MRI restoration across different acceleration factors.

Main Methods:

• Developed MR-IPT, a ViT-based image-domain framework.
• Pre-trained MR-IPT on a large-scale dataset with diverse undersampling patterns and acceleration settings.
• Leveraged a shared transformer backbone for universal feature representation learning.

Main Results:

• MR-IPT demonstrated superior performance compared to CNN-based and existing transformer-based methods.
• Achieved high-quality MRI restoration across various acceleration factors and sampling masks.
• Showcased strong robustness and maintained performance on unseen acquisition setups.

Conclusions:

• Transformer-based general models significantly advance MRI restoration.
• MR-IPT offers improved adaptability and stability over traditional deep learning approaches.
• The proposed framework presents a scalable and efficient solution for accelerated MRI.